How to Choose AI Accounting Software

How the software behaves when it is wrong or unsure — not how it behaves when it is right. Every product in this category demos beautifully, because demos use clean documents from familiar suppliers, and any competent system handles those. The differences only appear at the edges: an unfamiliar supplier, a badly photographed receipt, a bill that duplicates one from last week through a different channel, a transaction that could plausibly match two entries. A good system flags those with the reason attached and refuses to guess. A bad one produces a confident, plausible, wrong answer — which is far more dangerous than an obvious failure, because it passes review. So construct your evaluation around your ugliest documents, not your tidiest, and watch specifically for what lands in the exception queue. A vendor whose exception queue is suspiciously empty is not being clever; they are hiding uncertainty from you.

Ask what the number counted, and watch what happens next. "99% accurate" is unfalsifiable without its denominator: field-level accuracy on clean native PDFs from a fixed supplier set is an easy test that predicts nothing about your intake, while document-level accuracy across a real firm’s mixed pile is the measurement that matters and is far less flattering. The follow-up question is better still: what proportion of documents go straight through without human intervention, on files like ours? If the vendor cannot answer, or answers with a number that does not depend on client mix and document quality, they are quoting marketing rather than measuring. The only trustworthy evidence is a trial on your own documents, including the bad ones. Any vendor confident in their product will agree to that immediately.

Four things, in writing. Where is client data stored and processed, and in which jurisdiction — you are handling other people’s financial records and your obligations under the Privacy Act 1988 do not travel with the data. Second, is one client’s data isolated from another’s structurally, or only by policy? For a firm this is not a nice-to-have; a tool that could surface one client’s information in another’s context is unusable regardless of how well it codes. Third, is your client data used to train models that serve other customers, and can you opt out? Fourth, what happens to the data if you leave — what you can export, what is deleted, and when. Vague answers to any of these are themselves the answer.

Give it a supplier it has never seen, with a layout unlike anything in its samples. Template-based OCR wearing an AI label falls apart immediately, because it was matching field positions rather than understanding a document — that is why those products need a template per supplier and break when a supplier redesigns their invoice. Then give it a document that is internally inconsistent: line items that do not sum to the total, GST that does not reconcile. A reading tool extracts the numbers and passes them on. A system that understands the document notices the contradiction and flags it. Finally, ask what happens when the same invoice arrives twice through different channels a week apart. Duplicate detection based on genuine understanding catches it; character recognition does not.

Not for its own sake, but there are three specific reasons it often matters, and it is worth being precise about them rather than patriotic. First, the domain: BAS, GST treatment, the ATO’s expectations, TPB obligations and the shape of an Australian chart of accounts are not universal, and a product built for another market handles them as an afterthought — which shows up in exactly the edge cases that cost you. Second, data jurisdiction, which is a live question under the Privacy Act 1988 when you are custodian of other people’s financial records. Third, support at the times you actually need it: BAS week is not a good moment to discover your vendor’s support desk wakes up when your quarter is already lodged. If an overseas product wins on all three, buy it — but check all three.

A great deal, unprompted — and the absence of that conversation is itself diagnostic. Any honest vendor in this category should volunteer that the AI does not exercise professional judgement, does not sign off on figures, does not lodge with the ATO, and does not replace the accountability of a registered agent. They should tell you which of your files are poor candidates and why. They should tell you if what you are struggling with is already solved by software you own. A vendor who has no limits to describe has either not thought about the problem or is not telling you what they found. In our own case the list is on nearly every page of this site, because a firm that discovers the boundaries after purchase discovers them at the worst possible moment.